auto merge of #14296 : kballard/rust/diagnostic_color_newline, r=alexcrichton

When printing colored diagnostics, we need to reset the terminal before
emitting the newline, not after. Otherwise it gets line-buffered and the
color won't reset until the next line is printed or the compiler exits.

Normally this isn't a problem, but when running rustc in parallel with
other processes (e.g. `make -j4`) this can cause the color to leak
to other lines.

Author: bors

src/libsyntax/diagnostic.rs

@@ -232,8 +232,27 @@ fn print_maybe_styled(w: &mut EmitterWriter,
     match w.dst {
         Terminal(ref mut t) => {
             try!(t.attr(color));
-            try!(t.write_str(msg));
-            try!(t.reset());
+            // If `msg` ends in a newline, we need to reset the color before
+            // the newline. We're making the assumption that we end up writing
+            // to a `LineBufferedWriter`, which means that emitting the reset
+            // after the newline ends up buffering the reset until we print
+            // another line or exit. Buffering the reset is a problem if we're
+            // sharing the terminal with any other programs (e.g. other rustc
+            // instances via `make -jN`).
+            //
+            // Note that if `msg` contains any internal newlines, this will
+            // result in the `LineBufferedWriter` flushing twice instead of
+            // once, which still leaves the opportunity for interleaved output
+            // to be miscolored. We assume this is rare enough that we don't
+            // have to worry about it.
+            if msg.ends_with("\n") {
+                try!(t.write_str(msg.slice_to(msg.len()-1)));
+                try!(t.reset());
+                try!(t.write_str("\n"));
+            } else {
+                try!(t.write_str(msg));
+                try!(t.reset());
+            }
             Ok(())
         }
         Raw(ref mut w) => {